On Mon, 2 Nov 2020 15:58:15 GMT, Coleen Phillimore <[email protected]> wrote:
>> This change turns the HashTable that JVMTI uses for object tagging into a
>> regular Hotspot hashtable - the one in hashtable.hpp with resizing and
>> rehashing. Instead of pointing directly to oops so that GC has to walk the
>> table to follow oops and then to rehash the table, this table points to
>> WeakHandle. GC walks the backing OopStorages concurrently.
>>
>> The hash function for the table is a hash of the lower 32 bits of the
>> address. A flag is set during GC (gc_notification if in a safepoint, and
>> through a call to JvmtiTagMap::needs_processing()) so that the table is
>> rehashed at the next use.
>>
>> The gc_notification mechanism of weak oop processing is used to notify Jvmti
>> to post ObjectFree events. In concurrent GCs there can be a window of time
>> between weak oop marking where the oop is unmarked, so dead (the phantom
>> load in peek returns NULL) but the gc_notification hasn't been done yet. In
>> this window, a heap walk or GetObjectsWithTags call would not find an object
>> before the ObjectFree event is posted. This is dealt with in two ways:
>>
>> 1. In the Heap walk, there's an unconditional table walk to post events if
>> events are needed to post.
>> 2. For GetObjectWithTags, if a dead oop is found in the table and posting is
>> required, we use the VM thread to post the event.
>>
>> Event posting cannot be done in a JavaThread because the posting needs to be
>> done while holding the table lock, so that the JvmtiEnv state doesn't change
>> before posting is done. ObjectFree callbacks are limited in what they can
>> do as per the JVMTI Specification. The allowed callbacks to the VM already
>> have code to allow NonJava threads.
>>
>> To avoid rehashing, I also tried to use object->identity_hash() but this
>> breaks because entries can be added to the table during heapwalk, where the
>> objects use marking. The starting markWord is saved and restored. Adding a
>> hashcode during this operation makes restoring the former markWord (locked,
>> inflated, etc) too complicated. Plus we don't want all these objects to
>> have hashcodes because locking operations after tagging would have to always
>> use inflated locks.
>>
>> Much of this change is to remove serial weak oop processing for the
>> weakProcessor, ZGC and Shenandoah. The GCs have been stress tested with
>> jvmti code.
>>
>> It has also been tested with tier1-6.
>>
>> Thank you to Stefan, Erik and Kim for their help with this change.
>
> Coleen Phillimore has updated the pull request incrementally with one
> additional commit since the last revision:
>
> Code review comments from StefanK.
Looks great in general. Great work Coleen, and thanks again for fixing this. I
like all the red lines in the GC code. I added a few nits/questions.
test/hotspot/jtreg/vmTestbase/nsk/jvmti/scenarios/capability/CM02/cm02t001/cm02t001.cpp
line 656:
> 654: result = NSK_FALSE;
> 655:
> 656: printf("Object free events %d\n", ObjectFreeEventsCount);
Is this old debug info you forgot to remove? Other code seems to use
NSK_DISPLAY macros instead.
src/hotspot/share/prims/jvmtiTagMap.cpp line 345:
> 343:
> 344: // Check if we have to process for concurrent GCs.
> 345: check_hashmap(false);
Maybe add a comment stating the parameter name, as was done in other callsites
for check_hashmap.
src/hotspot/share/prims/jvmtiTagMap.cpp line 3009:
> 3007: // Lock each hashmap from concurrent posting and cleaning
> 3008: MutexLocker ml(tag_map->lock(),
> Mutex::_no_safepoint_check_flag);
> 3009: tag_map->hashmap()->unlink_and_post(tag_map->env());
This could call unlink_and_post_locked instead of manually locking.
-------------
Changes requested by eosterlund (Reviewer).
PR: https://git.openjdk.java.net/jdk/pull/967
